ReViT
收藏资源简介:
ReViT数据集是一个专为旋转等变视觉变换器(ReViT)项目设计的基准数据集,用于学习和预测偏微分方程描述的物理系统动态。它包含三个高分辨率数值模拟子集:KF2D(2D Kolmogorov流,模拟二维湍流,数据包含速度场,空间分辨率160×160,训练集50条轨迹各51时间步,测试集30条轨迹各201时间步)、MHD_64(3D磁流体动力学,模拟可压缩理想MHD,数据包含7个物理场如密度和磁场,空间分辨率64×64×64,5条轨迹各100时间步)和P3D(3D周期性通道流,模拟不可压缩湍流,数据包含速度场和压力场,空间分辨率96×96×96,训练数据1条模拟180时间步,测试数据20时间步)。数据集总大小在10GB到100GB之间,采用MIT许可证,旨在为开发旋转等变机器学习模型提供标准化基准,适用于物理信息机器学习、时间序列预测和PDE系统动态研究。
The ReViT dataset is a benchmark dataset specifically designed for the Rotation-Equivariant Vision Transformer (ReViT) project, which is utilized to learn and predict the dynamics of physical systems governed by partial differential equations (PDEs). It includes three high-resolution numerical simulation subsets: 1. KF2D (2D Kolmogorov Flow): simulates two-dimensional turbulence, containing velocity field data with a spatial resolution of 160×160. The training set consists of 50 trajectories each with 51 time steps, while the test set has 30 trajectories each with 201 time steps. 2. MHD_64 (3D Magnetohydrodynamics): simulates compressible ideal magnetohydrodynamics, encompassing 7 physical fields such as density and magnetic field, with a spatial resolution of 64×64×64 and 5 trajectories each having 100 time steps. 3. P3D (3D Periodic Channel Flow): simulates incompressible turbulence, including velocity field and pressure field data with a spatial resolution of 96×96×96. The training data is 1 simulation run with 180 time steps, and the test data has 20 time steps. The total size of the dataset ranges from 10 GB to 100 GB, and it is released under the MIT License. This dataset aims to provide a standardized benchmark for developing rotation-equivariant machine learning models, and is applicable to research in physics-informed machine learning, time series prediction, and PDE system dynamics.




